KS150/dpo-qwen-cot-merged

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Feb 4, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

The KS150/dpo-qwen-cot-merged model is a fine-tuned variant of the Qwen3-4B-Instruct-2507 architecture, optimized using Direct Preference Optimization (DPO) via the Unsloth library. This 4 billion parameter model focuses on enhancing reasoning capabilities through Chain-of-Thought (CoT) and improving structured response quality. It is designed for applications requiring aligned, high-quality outputs, particularly in tasks benefiting from improved logical progression and structured answers.

Loading preview...

Model Overview

KS150/dpo-qwen-cot-merged is a specialized language model derived from the Qwen3-4B-Instruct-2507 base model. It has undergone Direct Preference Optimization (DPO) using the Unsloth library, aiming to align its responses more closely with preferred outputs.

Key Capabilities

  • Enhanced Reasoning: Optimized to improve Chain-of-Thought (CoT) reasoning, leading to more logical and coherent multi-step responses.
  • Structured Output Quality: Fine-tuned to produce higher quality and more structured responses based on preference datasets.
  • Full-Merged Weights: Distributed as full-merged 16-bit weights, eliminating the need for adapter loading and simplifying deployment with transformers.

Training Details

The model was trained for 5 epochs with a learning rate of 5e-04 and a maximum sequence length of 1024. The DPO process utilized a beta value of 0.1. The underlying LoRA configuration (r=8, alpha=16) has been merged into the base model.

Ideal Use Cases

This model is particularly well-suited for applications where response quality, logical reasoning, and structured output are critical. Developers can leverage its enhanced alignment for tasks requiring precise and well-reasoned answers, especially those benefiting from Chain-of-Thought prompting.